It was the wee hours of Saturday morning when I got on a zoom call to advise a founder on an AI product. A 30 min discussion on details and the engineering team promised a PRD of their ideas to be delivered the next day (side note - I wonder why people still write PRDs when you can have demos in approx. same time). Next day, I received the document and I immediately lost all motivation to read it.
You know how it feels to be handed a 10 page AI generated word salad. I mean, in the old world, PRDs already felt like strong sedatives. But that part aside, the core question that started forming in my mind while I turned glassy eyed reading the em-dashes and the “it’s not this, it’s that!” punch lines was - how do I trust this document?
This question is deceptively hard and incredibly important.
It’s important because I care about this product and you want it handled carefully. It’s one thing to let Claude run wild with vibes on personal projects, but here we’re talking about work that will go out into the world, that humans will use and trust their data with. Most importantly a thing I will measure myself by. So it’s important that I trust the plan, the output, the underlying machinery before I step on it.
It’s also deceptively hard because remember LLMs are great at next token prediction and literally trained to emit human language, do reasoning and sound sensible. Reading this document I would’ve totally convinced myself it’s high quality with reasonable likelihood. But have you ever tried asking an LLM to critique its own plan? It will give you another long document full of gaps and that will sound convincing too! “Convincing” is not trustworthy. This is also why I couldn’t just give the AI slop to my Claude and ask it to generate another equally sloppy counter-narrative.
Trust matters to people who care about things.
The way to generate trust is to create ownership. Ownership implies responsibility. By a human. THE human who ships the output. It’s basically saying - “do you sign off on every single line of output you have generated and put your reputation at stake on it?” - every time someone generates output with AI.
I know it’s hardcore and unnecessary to take this oath every time you ship something with AI, but it changes something fundamental in the way you ship. You would either double check your own generated output which would be a massive win for the maintainability crisis, or you would invest in building AI systems for yourself that are trustworthy in design so that you don’t have to double check each time. Both, IMO, are big wins making the world a better place.
Let’s zoom out for a moment and truly imagine the scale of this problem. Imagine how many times you’re slapped with an AI word-salad to climb out of - and how many times do you actually fight it ground up? Imagine all the vibe-code being written into production databases - directly impacting reliability of even the biggest enterprise SaaS companies. It’s hard to tell slop from not cuz proving not slop is a steeply high effort vs generating slop is cheap and instantaneous.
This is a deep problem in the world of AI engineering too. There are terms for it - verifiers, evals, benchmarks. All this work to basically say - “look here’s proof why you should trust this output of this AI system”. Benchmarks are known ground truths that LLMs are run against to prove trust on various tasks. Evals are frameworks developers deploy, with humans-in-the-loop to prove the AI is doing well with people interacting with it. Then there’s the problem of verifiers - AI to check another AI’s work - the holy grail that unblocks AI factories for software, content, science, drug-development, everything - and yet unsolved. Proof matters because it generates trust.
Addy Osmani, in his keynote at AIE 2026, touched on this exact topic from a different angle - what should software engineers be doing in a world where coding is AI-automated? His answer - become owners of things because it’s a defensible moat to represent work even if it’s AI generated. The unfinished part of the thought was - you can only defend what you trust - so really, become owners who are trustworthy.
Trust matters. People work hard to create trust repetitively by ownership aka reputation.
We need owners.